Skip to main content

Simulation run on this device. Scenario values are not sent anywhere — the engine is fully local.

ai simulation · free

Attention Atlas

Paint per-head patterns from a small attention trace and name the behaviors.

Which head is doing induction — and which is noise?

Engine 1.0.0 · simulation
464
16512
0.55

How peaked each head's mass is.

Mean concentration
0.46
Peak head
0.69
Induction-like heads
0 / 16
Top-8 mass (toy)
6%128 tokens

Diffuse attention — hard to name a head's job; look for structure before ablating.

seq=128 heads=16 focus=0.55
entropy proxy 54% · induction-like 0

Method

  • Synthetic per-head concentration scores vs sequence length; induction-like heads are the peaked ones under the focus knob.
  • Numbers are teaching approximations, not production capacity guarantees.